Retrieval

Vector Index

The vector index provides fast semantic candidate retrieval for paraphrase-heavy queries.

Embedding path

A bi-encoder converts memory text and queries into a shared dense space. Vector similarity retrieves semantically close passages even when wording differs.

Normalize vectors consistently for both ingest and query paths to avoid score drift.

HNSW tradeoffs

Higher `ef_search` improves recall but increases latency.
Higher `M` improves graph connectivity but uses more memory.
Batch insertion patterns influence graph quality and cold-start behavior.

Typical tuning

vector_index:
  metric: cosine
  m: 24
  ef_construction: 200
  ef_search_default: 64

Operational checks

Track recall@k on a fixed evaluation set.
Monitor p95 query latency by entity size bucket.
Verify vector-id to memory-id mapping consistency after restarts.